-
TrajAudit – Live trajectory audit dashboard built with Streamlit and Plotly
-
Interactive trajectory visualization with dynamic filtering
Inspiration
Every 3 minutes an Indian loses money to a scam bot. Current AI detects scams too late — after victims already pay. We built TrajAudit to catch scam intent early.
What it does
TrajAudit analyzes conversation trajectories to detect behavioral phase transitions in scam dialogues — Rapport → Extraction → Capture → Conversion. It flags scam bots at Turn 3–5, where LLM moderation only triggers at Turn 12–18. 100x cheaper than LLM-based moderation.
How we built it
Two-layer system: lightweight rule-based signal extractor (money requests, urgency language, false proximity detection) + Gemini AI confirmation on flagged turns only. Risk scores plotted as a trajectory graph using Plotly.
Core Insight
Scam bots have a hidden objective. That objective leaks through behavioral phase transitions even inside friendly small talk. TrajAudit detects the directional drift — not individual keywords.
What makes it novel
Most fraud detection flags keywords. TrajAudit tracks conversation trajectory — the direction of intent across turns. This is behavioral intelligence, not keyword filtering.
Challenges
Parsing real Snapchat conversation data and separating intentional scammer noise from genuine rapport patterns.
What we learned
Scam conversations follow deterministic phase progressions. The trajectory is the signal — not any single message.
Future work
Scale to 10,000+ labeled conversations, integrate with messaging platforms in real-time, publish as peer-reviewed behavioral fraud detection research.
Log in or sign up for Devpost to join the conversation.